Most AI agents these days have a focused narrow skill. For example you can make an AI agent that detects emotions in a video but most likely is not complete. That's OK, you can focus on improving those specific skills and you can still benefit from a wide variety of other skills provided AI agents in the network by joining to the network.
Create well defined standards and interfaces for boarcasting and discovering the type of skills and AI agent provides.
A trust system must be created to review the quality of AI agents provided in the network.
You can join by participaing on discussion we have on gitter.im/ark-ai.
Also you can contibute code on github.com/ark-ai.
The project is at the very beginnings and currenly we are currently defining the main objectives.
The next phase is to figure out what would be the best course of action to make this idea an reality.